OpenAI used its own AI models to help redesign parts of a chip. The people running the project can't fully explain what the AI did.
Richard Ho, who leads OpenAI's hardware team, said AI models were turned loose on portions of the company's Jalapeno chip design. In at least one case, the AI-driven changes cut die area by more than 13%. Ho told interviewer Ian Cutress, in an interview published September 30, that his engineers can't always reconstruct the reasoning behind those changes. The interview didn't specify which part of the chip was affected or how the AI arrived at its choices.
A 13% reduction in die area is a real number, not a demo flourish. On a custom AI accelerator, less area per chip means lower manufacturing cost and more dies per wafer. But it's also a tell: a hardware team is now shipping optimizations its own engineers can't fully audit, the same explainability problem that has dogged black-box AI software, now showing up in physical silicon.
Chip design has leaned on automation for decades, but when the person running the project still says he doesn't fully trust it, that's a useful corrective to any pitch that AI-designed chips are already a solved problem.